Does zero-sum mean the expected payoff is zero for each player?

No. The two values can be 0.5 and -0.5, as in the example.

Must players use randomization?

Not always. Some games have optimal pure strategies; others require mixed strategies to achieve the minimax value.

Are all competitive situations zero-sum?

No. Competition can coexist with shared benefits, shared costs, and changes in total value.

AI Glossary / Graphs and symbolic AI / Intermediate

Zero-sum games

Definition

Zero-sum games are games in which the players' payoffs add to zero for every outcome. In the common two-player case, one player's gain is exactly the other's loss under the chosen payoff model. This makes their objectives directly opposed and supports minimax reasoning: each player plans against the strongest response the other can make.

What it means

The zero refers to the sum of payoffs, not to a requirement that each player receives zero or that the game is evenly balanced. One player may have a structural advantage and a positive expected payoff under optimal play.
The term describes modeled utilities. A contest can be zero-sum in points while producing shared enjoyment or imposing shared costs outside that score. Constant-sum games have a closely related strategic structure after a suitable payoff transformation. Most business relationships should not be presumed zero-sum merely because competition exists; cooperation, costs, and changes in total value can alter both parties' outcomes.

How it works

For a finite two-player matrix game, specify the row player's payoff for each pair of actions. The column player's payoff is its negative. The row player maximizes expected payoff while the column player minimizes that same quantity.
A pure strategy chooses one action; a mixed strategy assigns probabilities to actions. The minimax theorem ensures that finite two-player zero-sum games have a value when mixed strategies are allowed: the best guaranteed expected payoff equals the opponent's corresponding best upper bound. This is an expectation-based guarantee under the model, not a promise about the result of every individual play.

Three practical examples

A board-game program can model a win as +1, a loss as -1, and a draw as 0 for one player, with opposite payoffs for the other.
A classroom matrix game can demonstrate why randomization prevents an opponent from exploiting a fixed action.
A controlled multiagent-learning experiment can use opposing rewards to test whether agents approach a known minimax benchmark.

Think of it this way

Think of moving points between two accounts in a closed scoring system: every point credited to one is debited from the other. The analogy is limited because real interactions may create benefits or costs outside those accounts, and the strategic problem also depends on available actions, information, and timing.

Worked example

Consider an original simultaneous two-player game. The row player chooses Top or Bottom; the column player chooses Left or Right. The row payoffs are: Top-Left gives 2, Top-Right gives -1, Bottom-Left gives -2, and Bottom-Right gives 3. The column payoff is always the negative of the listed number.
If the row player always chooses Top, the column player can choose Right and hold its payoff to -1. If it always chooses Bottom, Left gives it -2. The best pure-strategy guarantee is therefore -1.
Now let the row player choose Top with probability p. Against Left, expected payoff is 2p - 2(1-p) = 4p - 2. Against Right, it is -p + 3(1-p) = 3 - 4p. Equalizing these expressions gives 8p = 5, or p = 5/8. Both expected payoffs then equal 0.5.
For the column player, let q be the probability of Left. Top produces expected row payoff 2q - (1-q) = 3q - 1. Bottom produces -2q + 3(1-q) = 3 - 5q. Equalizing gives q = 1/2, with row payoff 0.5 in either case.
Together these strategies establish the game value: the row player can guarantee at least 0.5 in expectation, and the column player can prevent more than 0.5. The column player's value is -0.5, so the sum remains zero. Individual outcomes are still 2, -1, -2, or 3 for the row player; none is required to equal the expected value.

Illustrative case study

Fictional case: a game-simulation team trains two agents with rewards that are exact negatives. It notices that one agent beats its current training partner frequently and initially calls the policy unbeatable.
The team instead tests against several responses, including strategies designed to exploit predictable choices. It compares performance with the known value of a small benchmark game before extending the experiment. Beating one opponent is useful evidence about that matchup, but does not establish a minimax guarantee against all allowed opponents.

Common misconceptions and mistakes

Do not confuse zero-sum with fair, symmetric, deterministic, or perfect-information. Those are separate properties. A zero-sum game can involve hidden information or chance.
Do not interpret a mixed strategy as alternating predictably through actions. If the opponent can infer the next action, the intended simultaneous randomization model may no longer apply. Also distinguish the payoff model from the entire real-world situation: adding costs borne by both players can make their total payoffs vary, changing the zero-sum assumption.

Business use

Zero-sum models provide a clear framework for strictly opposed objectives in simulations and adversarial decision problems. They are valuable when the assumption accurately captures the question being studied.
Before using the model, specify the players, actions, information, timing, and payoff units. Check whether shared gains or losses have been omitted. For AI evaluation, measure how exploitable a policy is rather than relying only on wins against one fixed opponent. If the setting includes opportunities for mutual benefit, a general-sum model may better represent the actual choices.

Check your understanding

Why is the row player's 0.5 expected payoff compatible with a zero-sum game?

Answer: The column player's expected payoff is -0.5. The sum is zero even though neither individual expectation is zero.

Related terms

Game theory
Minimax
Mixed strategy
Nash equilibrium
Payoff matrix
Constant-sum game
General-sum game
Exploitability

Learn next

Change the row player's Top probability to one half while keeping the same payoff matrix. Its expected payoff against Left becomes 0 and against Right becomes 1. Explain why an opposing column player chooses Left and why this is worse than the 0.5 guarantee at p = 5/8.

Sources

Primary sources: AIMA original two-player game context: https://raw.githubusercontent.com/aimacode/aima-glossary/master/sentences.txt ; Daskalakis and Farina, MIT original learning-in-games notes and minimax theorem: https://www.mit.edu/~6.7980/learning_intro.html . The payoff matrix and equilibrium arithmetic are original.

Previous: Zero-shot TTS Next: Zero-width